Open-source AI agents for companies in Germany

What open-source AI agents are

An AI agent is a language model that works through a task with tools: it reads documents and calls other systems until the task is done. With open-source agents, the software around the model is open, from the control plane that assigns the work to the harness that runs each agent. You can read that code and run it on your own server.

Paperclip is such a control plane, under the MIT license. It keeps the org chart, the goals, the issues and the approvals, and it wakes an agent with a heartbeat when work waits. It runs no model itself. Each agent runs through an adapter, for Claude, Codex, OpenCode, Pi or a plain script.

Why a company runs agents on its own server

The control plane holds your tasks and the record of each approval, so it belongs on a server you control. The model is a separate decision: the agents call a cloud model through its API, or an open-weight model that runs on your hardware.

Open code shows what the software does. It does not stop an agent from doing harm. On January 28, 2026, Cisco reported a third-party skill for the open-source agent OpenClaw that sent data to an outside server and used prompt injection, and found at least one vulnerability in 26 percent of 31,000 agent skills it analyzed (Cisco). An open-source agent needs the same limits as any other: the rights of a new colleague, and a person who confirms changes to live data.

How I set up a company of agents

For a client’s information platform I set up part of the work like a company whose staff are AI agents: a CEO agent, a chief each for data, technology and marketing, and an engineer under the technology chief. The first jobs produced the rules I apply to agents today:

• A chief never executes. It hands each task to an idle agent of its team or creates a new one, so no agent is booked twice.
• One job is in flight per agent. Two jobs of one agent in one workspace can destroy each other’s work.
• A run and a task have separate clocks. The control plane ends a run after about four minutes, so long jobs run detached and get polled, or they become child issues.
• A report gets measured. In the first job an agent did the work right and reported the wrong count, because it counted a wider area than the task named. Since then a verifier script measures each result.
• Confirmations reach a phone. I put an OAuth gateway, built with FastMCP, in front of the control plane’s MCP server, so Claude on a phone reads and writes the issues.

The full account is in the post Running a company of AI agents with Paperclip.

Checklist before an agent takes its first job

Questions to answer for each agent in your company. Bring your answers to our call.

1. Which role does the agent have, and which chief assigns its work?
2. Which systems may it read, and which may it write?
3. Does it run against a development system only?
4. Who confirms a change to live data, and through which channel outside the chat?
5. Which script checks its reports, independent of the agent?
6. What is its monthly budget?
7. How does a job continue that takes longer than one run?
8. Which model does it call, and where does that model run?
9. Who may pause it, and who reads its history afterwards?

Who I am

I am Michael Wutzke, an AI engineer in Frankfurt with more than twenty years in IT and media. I was Head of Decentralized Finance and Node Operations, then CIO, at the Frankfurt-based company Blocksize Capital. Today I help a client rebuild its platform with Claude and Codex; the work has produced more than 300 API endpoints and more than 8,000 commits. At Claude Hacker House I teach building MCP servers and virtual organizations of agents. Details: Career stages.